ollymcbride4/docdrift-bugs

DocDrift bugs: docs-vs-code drifts in 12 Python libraries, found with Beam

DocDrift gives Beam a whole Python library (docs + code, up to ~400k tokens) in one call. A free AST symbol check proposes name and signature mismatches. Beam triages those candidates against the packed repo, can add cross-file or prose drift, and writes a script that runs the documented usage. A finding counts as reproduced only if that script raises TypeError, AttributeError, ImportError, or NameError on a documented-usage line.

This dataset is the recorded output of that pipeline on 12 pinned libraries. 34 real docs bugs were found in 11 of the 12 libraries; 30 reproduce. Labels were checked against the source by an AI reviewer, not a human annotator. The stored field is still named hand_verdict.

Licence. DocDrift's own labels, verdicts, and notes are CC BY 4.0. Quoted documentation and source excerpts remain under each library's own licence (all 12 are permissive OSS).

from datasets import load_dataset
ds = load_dataset("ollymcbride4/docdrift-bugs", "findings")

• Space: https://huggingface.co/spaces/ollymcbride4/docdrift
• Demo video: https://huggingface.co/spaces/ollymcbride4/docdrift/resolve/main/assets/docdrift-demo.mp4
• Code: https://github.com/OlsOls4/repoatlas-1m
• Dataset: https://huggingface.co/datasets/ollymcbride4/docdrift-bugs
• HF thread: https://huggingface.co/reflection/Beam-1.0-501B-A23B/discussions/5

How it was built

1. Pin each library to a recorded commit.
2. Pack docs + source up to a 400k token cap.
3. Run a free symbol check, then one whole-repo Beam triage call per library (11 of 12 succeeded in one call; structlog failed; cattrs needed 2).
4. Beam-written repro scripts, plus one hand-written starlette script from Beam's verdict.
5. Citation check is quote alignment of cited lines vs packed source (±3-line shift). It is not a verdict that the drift is real.

Configs

config	rows	what it is	fields
findings	60	Hybrid method items (the 60 findings)	repo, id, kind, verdict, drift_id, beam, strict_confirmed, repro_status, evidence, hand_verdict, hand_verdict_description
triage_verdicts	56	Beam triage verdicts (56 judged)	repo, candidate_id, source_id, verdict, reason, code_path, code_line, code_quote
repros	54	Executable repro cases	case_id, repo, status, strict_confirmed, exception_type, error_excerpt, evidence
planted_drift	58	Planted-drift set + ground truth	original probe fields plus seed
ablation_verdicts	56	Whole-repo vs local-file vs matched-cost retrieval triage on the same 56 candidates	key, repo, candidate_id, label, hand_verdict, hand_verdict_description, cross_file, gold_outside_local, whole_verdict, local_verdict, matched_verdict, stability_verdict, citation flags
matched_cost_ext_verdicts	63	Scored pre-registered extension items only (63 rows in this file; 49 cross-file, 48 outside-local). Pooled 52/78 includes the original 56 from ablation_verdicts; those original rows are not duplicated here.	key, set, stratum, label, hand_verdict, cross_file, gold_outside_local, per-arm whole_ / local_ / matched_ verdicts and citations
beamv3_findings	115	Extra 32k-docstring-chunk and docs-pass findings (unlabelled not counted as true)	id, repo, pass, doc_path, doc_line, label, drift_id, label_by, hand_verdict, hand_verdict_description, new_vs_hybrid, strict_confirmed, explanation
token_ledger	493	Hybrid + first-sweep + ablation + extra-pass token ledger	original ledger fields plus ledger (first_sweep, hybrid, ablation, beamv3, matched_cost, matched_cost_ext, fresh_run)

hand_verdict means checked against the source by an AI reviewer. It is not a human annotator.

Limits

• The 45-drift universe is the union of method output, so it favours those methods. The 45 true hybrid findings are a different set from those 45 pooled drifts.
• 11 of 12 libraries were triaged; structlog failed; cattrs needed 2 calls.
• 2 false + 1 unclear received a strict badge (werkzeug S38, S39 false; S41 unclear).
• End-to-end including repro is 9,302,105 (~9.3M) = 4,424,575 first sweep + 4,739,392 hybrid triage + 138,138 repro. The first sweep is in token_ledger as ledger=first_sweep (12 rows from the recorded 400k traces).
• Extra experimental runs, not in the 9.3M: ablation 1,948,469; beam v3 7,713,453; matched-cost control 3,957,434; matched-cost extension 7,528,878; tomlkit CLI run 185,842. Matched-cost controls plus the recording: 11,672,154.
• Long context did not change verdict accuracy on these 56 candidates; it changed citation grounding. A matched-cost retrieval control (96.7% of whole-repo spend) matched accuracy (46/56 vs 47/56) but not grounding (cross-file 19/30 vs 25/30; McNemar p = 0.070, not shown at this n). Retrieval missed the defining file in 5 of 13 outside-local items. A pre-registered extension added 63 items (57 planted renames, 6 real structlog). Pooled, retrieval grounded 38/78 cross-file cases vs 52/78 for the whole repo (McNemar p = 0.0094). Outside-local: 38/61 vs 25/61 (p = 0.0146). Accuracy still did not differ (105/118 vs 108/119). The new items alone give p = 0.096 (not shown at this n), and the pooled p is 0.18 without one pre-registered parse-failure re-send. Pre-registration: design_ext.md sha256 42f6b6fe…c996.
• 32k docstring chunks recovered 7 of 11 hybrid misses; the docs pass recovered 1. New-item precision is a lower bound (18.5% for hybrid + 32k chunks). Headline stays 34 of 45.

Citation

DocDrift bugs: docs-vs-code drifts in 12 Python libraries, found with Beam. Labels by an AI reviewer against source at the pinned commit. Dataset: https://huggingface.co/datasets/ollymcbride4/docdrift-bugs